98% Accurate, Except It Wasn't: The FTC's Case Against an AI Detection Tool

The Federal Trade Commission gave final approval in August 2025 to an order against Workado, LLC, formerly known as Content at Scale AI, resolving allegations that the company falsely advertised its AI content-detection tool as "98 percent accurate" at determining whether text had been generated using artificial intelligence.DOCUMENTED

The FTC's own testing found the tool's actual accuracy fell to roughly 53% when applied outside the narrow category of text it had primarily been trained on, despite the company's marketing presenting the 98% figure as broadly applicable across general use cases.DOCUMENTED

An accuracy claim built on a narrow test set

According to the FTC, Workado marketed its AI content-detection tool as capable of accurately distinguishing AI-generated text from human-written text 98% of the time, a figure the company presented in its advertising without qualification as to what kind of text the claim applied to. The Commission's investigation found that the underlying model had in fact been trained largely on academic writing, a category of text with distinctive structural and stylistic patterns that do not necessarily generalize to the much broader range of writing styles found across everyday internet content, business communications, creative writing, or other general-purpose text.DOCUMENTED

When the FTC conducted its own testing of the tool's performance outside that narrow academic-text training set, the agency found accuracy fell to approximately 53% — only marginally better than a random guess between two categories, and a figure dramatically below the 98% claim used in the company's marketing.DOCUMENTED

Why the gap mattered

AI content detectors like Workado's tool are frequently marketed to educators, publishers, employers, and content platforms as a way to verify whether a given piece of writing was produced by a human or generated using an AI language model — a determination that can carry significant consequences, including academic integrity findings against students or content-moderation decisions on publishing platforms. A detector whose real-world accuracy falls dramatically short of its advertised claim risks generating false accusations against genuine human writers, particularly for text falling outside the narrow category the underlying model was actually trained to evaluate reliably.REVIEWED

The order's terms

The final order requires Workado to cease making any claims about the accuracy or efficacy of its AI content-detection tool, or any similar product, unless those claims are substantiated by competent and reliable evidence reflecting the tool's actual performance across the range of contexts in which it is marketed for use. The order does not appear to include a separate monetary penalty distinct from the injunctive relief governing the company's future advertising claims.DOCUMENTED

Part of a broader AI-claims enforcement push

The Workado case arrived as part of a series of FTC actions in 2025 targeting exaggerated or unsubstantiated claims about artificial intelligence products, an enforcement theme the agency has pursued using its ordinary consumer-protection authority rather than any AI-specific statute. FTC officials have repeatedly emphasized that adding the label "AI" to a product's marketing does not exempt a company from the same substantiation requirements that apply to any other performance or accuracy claim.DOCUMENTED

The case followed the agency's 2024 guidance outlining specific practices it considers off-limits for AI-related marketing, including misrepresenting what an AI product actually is or does, and making unsubstantiated claims about an AI tool's capabilities. The Workado order, together with the FTC's separate cases against AI-branded business opportunities and its broader inquiry into AI companion chatbots, reflects a consistent enforcement pattern: whatever the underlying technology, a specific, quantified performance claim like "98 percent accurate" must be backed by evidence covering the actual range of conditions under which the product is marketed for use, not just the narrow conditions under which it happens to perform best.REVIEWED

For consumers and institutions relying on AI-detection tools to make consequential decisions, the case is a reminder that a headline accuracy figure alone reveals little about a tool's actual reliability without knowing the specific conditions, and specific kind of text, under which that figure was measured.

The Workado case also illustrates a specific pitfall in how AI performance metrics are often communicated to the public: a single headline accuracy figure, without disclosure of the specific dataset or conditions under which it was measured, can create a badly misleading impression of a tool's real-world reliability. Educators, employers, and platforms relying on AI-detection tools to make consequential decisions about individual students or content creators may have little independent ability to verify a vendor's accuracy claims without the kind of detailed testing the FTC itself conducted in this case, underscoring the practical stakes of the agency's substantiation requirement.REVIEWED

The case also arrives at a moment when AI detection tools of this kind are being adopted rapidly across schools and universities specifically to address concerns about AI-assisted cheating, meaning the practical stakes of an overstated accuracy claim extend well beyond ordinary commercial disappointment into decisions that can carry academic and disciplinary consequences for individual students wrongly flagged by an unreliable tool.REVIEWED

The case is also notable for what it reveals about how AI marketing claims can fail even without any allegation of outright fabrication: Workado's tool did, in fact, perform close to its advertised accuracy on the narrow academic-text dataset it had actually been tested against, meaning the underlying deception lay not in inventing a number from nothing but in generalizing a narrow result to a much broader marketing claim the underlying testing never actually supported.REVIEWED

Have documents relevant to this story? Reach us through our tips channel.